/cmo ↔ mktg-studio integration

mktg-studio is the local dashboard that visualises everything you do. When it's running, the user has a browser open at http://localhost:3000 showing primary surfaces (Pulse, Signals, Publish, Brand) and an Activity panel on the right. Trend radar lives inside Signals. Audience summary and recommended next actions live inside Pulse.

This skill teaches you to drive that dashboard without being asked. No SDKs. No special tools. Every action is a curl you already know how to write.

1. Detect first

On every turn, before doing anything else marketing-y, run:

STUDIO=${STUDIO_BASE:-http://localhost:3001}
curl -fsS -m 2 "$STUDIO/api/health" >/dev/null 2>&1 && STUDIO_UP=1 || STUDIO_UP=0

• STUDIO_UP=1 → studio is up, follow the rules below
• STUDIO_UP=0 → studio is down, behave exactly as you would in a

CLI-only setup. Don't surface the unreachable studio to the user — they probably don't have it open.

Cache the result for the rest of the turn — don't hammer health.

If you want to confirm the wire shapes you're about to send, fetch the schema once per session (it's static):

curl -fsS "$STUDIO/api/schema" | jq '.routes[].path'
curl -fsS "$STUDIO/api/schema?route=/api/activity/log" | jq
curl -fsS "$STUDIO/api/help" | jq '.data.entryPoints'

2. The five verbs

Every meaningful action you take should map to one or more of these endpoints. Skip them when they don't apply — don't be noisy.

Copy-paste recipes

# Log a skill run
curl -fsS -X POST "$STUDIO/api/activity/log" \
  -H 'content-type: application/json' \
  -d '{"kind":"skill-run","skill":"landscape-scan","summary":"Refreshed landscape, 5 new claims","filesChanged":["brand/landscape.md"]}'

# Note a brand-file write (excerpt is the first ~800 chars)
EXCERPT=$(head -c 800 brand/voice-profile.md | jq -Rs .)
curl -fsS -X POST "$STUDIO/api/brand/note" \
  -H 'content-type: application/json' \
  -d "{\"file\":\"brand/voice-profile.md\",\"excerpt\":$EXCERPT}"

# Push an opportunity
curl -fsS -X POST "$STUDIO/api/opportunities/push" \
  -H 'content-type: application/json' \
  -d '{"skill":"keyword-research","reason":"You have a landscape but no keyword-plan yet — that gates seo-content","priority":80}'

# Toast
curl -fsS -X POST "$STUDIO/api/toast" \
  -H 'content-type: application/json' \
  -d '{"level":"success","message":"Landscape refreshed — see Trend radar"}'

# Navigate
curl -fsS -X POST "$STUDIO/api/navigate" \
  -H 'content-type: application/json' \
  -d '{"tab":"signals","filter":{"mode":"radar"}}'

3. Brand files are first-class

Writing a brand file directly with the Write tool works but doesn't update the studio. Prefer the HTTP path:

# 1. Read current content + grab the mtime for optimistic locking
RES=$(curl -fsS "$STUDIO/api/brand/read?file=voice-profile.md")
MTIME=$(echo "$RES" | jq -r '.data.mtime')

# 2. Build new content in-memory (your editing logic here)
CONTENT=$(jq -Rs . < new-voice-profile.md)

# 3. Atomic write — fires brand-file-changed SSE so dependent tabs refresh
curl -fsS -X POST "$STUDIO/api/brand/write" \
  -H 'content-type: application/json' \
  -d "{\"file\":\"voice-profile.md\",\"content\":$CONTENT,\"expectedMtime\":\"$MTIME\"}"

# 4. Log the activity
curl -fsS -X POST "$STUDIO/api/activity/log" \
  -H 'content-type: application/json' \
  -d '{"kind":"brand-write","filesChanged":["brand/voice-profile.md"],"summary":"Made the voice more casual — 3 anchor sentences swapped"}'

If the user edited the same file in their text editor while you were writing, the expectedMtime mismatch returns 409 CONFLICT with a fix hint pointing back to /api/brand/read to re-fetch + merge.

If you need to re-research a brand file, queue the owning skill:

curl -fsS -X POST "$STUDIO/api/brand/regenerate" \
  -H 'content-type: application/json' \
  -d '{"file":"landscape.md"}'
# → { ok: true, data: { jobId, skill, file } }
# poll via curl -fsS "$STUDIO/api/jobs/<jobId>" or watch /api/events

4. The idiomatic turn

A typical /cmo turn that runs a skill, writes a brand file, and recommends a next action:

STUDIO=${STUDIO_BASE:-http://localhost:3001}
curl -fsS -m 2 "$STUDIO/api/health" >/dev/null || exit 0   # 1. detect

curl -fsS -X POST "$STUDIO/api/activity/log" -H 'content-type: application/json' \
  -d '{"kind":"skill-run","skill":"landscape-scan","summary":"Refreshing landscape from 18 sources"}'

# 2. (do the actual work — Exa search, write brand/landscape.md)

EXCERPT=$(head -c 800 brand/landscape.md | jq -Rs .)
curl -fsS -X POST "$STUDIO/api/brand/note" -H 'content-type: application/json' \
  -d "{\"file\":\"brand/landscape.md\",\"excerpt\":$EXCERPT}"

curl -fsS -X POST "$STUDIO/api/activity/log" -H 'content-type: application/json' \
  -d '{"kind":"skill-run","skill":"landscape-scan","summary":"Done — 5 new claims, 2 deprecated","filesChanged":["brand/landscape.md"]}'

curl -fsS -X POST "$STUDIO/api/opportunities/push" -H 'content-type: application/json' \
  -d '{"skill":"competitor-alternatives","reason":"Fresh landscape — write 3 X vs Y pages","priority":75}'

curl -fsS -X POST "$STUDIO/api/navigate" -H 'content-type: application/json' \
  -d '{"tab":"signals","filter":{"mode":"radar"}}'

curl -fsS -X POST "$STUDIO/api/toast" -H 'content-type: application/json' \
  -d '{"level":"success","message":"Landscape refreshed — see Trend radar"}'

Eight calls. Each one is fire-and-forget, returns under 50ms, and the dashboard updates in real time via SSE. The user watches the work happen.

5. Don't break

• Every endpoint returns {ok: true, ...} or {ok: false, error: {code, message, fix?}}. Predictable envelope.
• Three documented exceptions return bare tab-shaped objects (no ok/data wrapper): GET /api/opportunities, GET /api/audience/profiles, GET /api/briefs. Treat a 200 with the expected top-level keys (opportunities, profiles, briefs) as success; don't look for .ok on these three.
• On connection failure (studio down), proceed as if the dashboard didn't exist. Don't retry.
• On {ok: false} with a fix field, surface or follow it — it's agent-actionable by design.
• /api/navigate and /api/toast are SSE-only (no DB write). Spamming them is harmless but ugly. Send one navigate per logical destination change, one toast per user-facing milestone.
• All mutating endpoints accept ?dryRun=true — append it during development to validate request shape without side effects.

6. What NOT to do

• Don't POST /api/navigate between every step of a multi-skill playbook — switch tabs once, at the end, when the user should look at the result.
• Don't POST /api/toast for routine logs — that's what /api/activity/log is for. Toasts are for things the user should drop everything and notice.
• Don't try to POST /api/brand/write outside brand/*.md — the path validator rejects everything else with BAD_INPUT.
• Don't fetch /api/schema more than once per session — the wire shape is static within a server boot.

Reference

• Full HTTP contract with copy-pasteable curls: docs/cmo-integration.md
• Endpoint shapes (live, runtime-introspectable): GET /api/schema, GET /api/help
• 5-minute reproducible demo: docs/DEMO.md